R-package decomposing tree-based model predictions (XGBoost, ranger, randomPlantedForest) into main effects and arbitrary-order interactions. From this decomposition one can read off exact interventional SHAP values, partial-dependence functions, and per-term variable importance.
Rust implementation of the FastPD algorithm with Python bindings — the same partial-dependence functions and functional ANOVA decompositions as the R glex package, but for XGBoost regressors and at much higher speed thanks to Rust code.
Tensor Separation Learning — a glass-box regression model that represents predictions as a sum of stages, each one a difference of two separable products of univariate functions. The factorization captures rich feature interactions while keeping every learned component directly inspectable through its 1D factors.